Everyone loves showing an agent that works — but what about an agent that survives, and learns something along the way? In this session we will violently kill an AI agent mid-orchestration, repeatedly, and watch it resume exactly where it left off. An agent's state lives in three places — in the conversation, in the execution, and in the knowledge — and each survives something different. We will meet Microsoft Agent Framework 1.0, the successor to Semantic Kernel and AutoGen, its Durable Task extension that adds durable execution to agent workflow graphs with automatic checkpointing, failure recovery, and human-in-the-loop approvals that can wait for days without consuming compute, and managed memory in Foundry Agent Service, with which we coach the agent once on a working procedure it then applies on its own in future runs — including procedural memory, inspection of stored records, and the right to be forgotten. We will explain why checkpoint storage alone is not durability, and how the very same workflow definition runs locally, on Azure Functions, or in your own container with zero code changes. All demos run in Python and are grounded in lessons from production agent systems in the insurance industry; this talk is for engineers moving agents from demo stage to systems that must also work at three in the morning.